Summary
What you’ll impact
Lead GPU Networking Engineer designs and implements networking primitives that connect thousands of GPUs into a high‑performance AI inference platform. Using C++ or Python, they integrate RDMA, RoCE, and InfiniBand, work with NCCL and NVSHMEM, and build low‑latency data paths, observability tools, and custom communication kernels to optimize distributed inference and wide‑expert parallelism. They collaborate with hardware and software teams to validate performance on cutting‑edge GPU clusters and enable sub‑second model startup.
Responsibilities
What you'll do
- Make RDMA First-Class: You will work on integrating RDMA/RoCE/InfiniBand capabilities directly into our inference stack, helping us move beyond TCP/IP to unlock order-of-magnitude improvements in bandwidth and latency.
- Optimize Distributed Inference: You will implement and tune the networking layers necessary for efficient Disaggregated KV Cache Offload and WideEP, ensuring seamless communication across NVLink and InfiniBand for our MoE models.
- Enable Serverless-Grade Startup Speeds for LLMs: You will work deeply with checkpointing and storage mechanisms to enable sub-10-second startup for trillion-parameter models.
- Deep-Dive into Hardware: You will characterize and validate networking performance on bleeding-edge clusters (H100/H200, B200/B300, GB200/300 NVL72), writing the acceptance tests that ensure our hardware delivers peak achievable throughput and minimal latency.
- Build Observability: You will design the tools that let us visualize packet flow, congestion, and effective bandwidth across the GPU interconnects, helping us diagnose complex distributed system behaviors.
- Optimize Kernels: You will work with communication libraries (NCCL, NVSHMEM) and potentially write custom communication kernels to overlap compute and data transfer.
Requirements
What you’ll bring
- You have deep experience with high-performance networking protocols (InfiniBand, RoCE v2) and understand the physics of data movement.
- You are fluent in C++ or Python, with the ability to bridge the gap between high-level logic and hardware. You have a deep understanding of the memory hierarchy in modern NVIDIA architectures (H100/Blackwell) and know how to optimize for it.
- You like going deep. You aren't afraid to dive into TensorRT-LLM source code, write custom C++ / Python bindings, or debug NVLink topology issues.
- You know when to use an off-the-shelf solution and when we need to build a custom solution because the upstream tools (like standard Kubernetes networking) are too slow for our needs.